The Core Hypothesis

As LLMs scale toward human-level performance, we are simultaneously becoming less capable of understanding how they produce their outputs. The central hypothesis of mechanistic interpretability (MI) is this: the internal computations of transformer-based LLMs are not random or inscrutable — they are structured, decomposable into interpretable circuits, and amenable to systematic reverse-engineering. If correct, this opens a credible path toward AI alignment through genuine understanding, not behavioral constraints alone.

Current State: Circuits, Superposition, and 2024–2026 Findings

Circuits: The Functional Units

The circuits framework proposes that specific groups of attention heads and MLP neurons collaborate to implement identifiable algorithms. By 2024–2025, this moved from theory to demonstrated reality:

• Induction heads — Attention heads implementing "copy previous token if context matches" underlie in-context learning (Olsson et al., 2022).
• IOI circuit — The Indirect Object Identification circuit in GPT-2 small was fully reverse-engineered: ~28 attention heads collaborating to complete sentences like "John gave Mary the book; Mary gave it back to ___." (Wang et al., 2022)
• Arithmetic circuits — Transformers implement modular arithmetic via a "Fourier multiplication" algorithm, encoding discrete Fourier transforms in embedding space (Lee et al., 2023).
• Entity binding — Factual recall tasks route through identifiable MLP layers acting as key-value memories for factual associations.

By mid-2025, circuit analysis extended to 70B+ parameter models, revealing qualitatively similar circuit motifs across architectures — suggesting circuits may be universal computational primitives, not training artifacts.

The Superposition Hypothesis

The superposition hypothesis (Elhage et al., 2022) is perhaps the field's most consequential theoretical contribution: neurons in LLMs don't implement one-to-one feature mappings. Instead, a single neuron encodes multiple features simultaneously — analogous to compressed sensing. This explains why earlier neuron-level interpretability failed.

The sparse autoencoder (SAE) approach (2023–2025) addresses this by projecting residual stream activations into a higher-dimensional sparse space where features do correspond to monosemantic concepts:

• Anthropic's SAE analysis of Claude 3 Sonnet (2024) identified 34 million features with interpretable semantic content — from "names ending in '-burg'" to "political rhetoric" to "DNA sequences."
• A feature for the concept "Assistant" activates strongly before output generation; clamping it produced significant behavioral shifts — the first demonstration of feature-level behavioral control.
• DeepMind's SAE work on Gemini showed similar sparsity and semantic coherence, confirming cross-architecture generalizability.

2025–2026 Landmarks

1. Representation Engineering at scale — Concepts like "honesty," "refusal intent," and "uncertainty" have stable linear representations across model families, readable with >85% accuracy and writable via activation steering (building on Zou et al., 2023).
2. Attention sink resolution — The attention sink phenomenon mechanistically explained as a softmax normalization artifact; models successfully modified at the circuit level without retraining.
3. Polysemanticity phase transitions — Theoretical work (Anthropic, 2025) modeled when superposition emerges as a function of dataset structure; models trained on datasets with ~10x more features than neurons show sharp polysemanticity transitions.

Practical Applications for Alignment and Safety

Detecting Deception Before It Manifests

In 2024, Anthropic demonstrated that "sleeper agent" models (trained with backdoors to exhibit harmful behavior under triggers) showed detectable anomalies in activation space even when behavioral outputs appeared normal. MI tools could serve as a computational lie detector — not relying on the model's self-reports.

Surgical Fine-Tuning via Circuit Targeting

MI-informed editing (ROME, Meng et al. 2022; MEMIT and successors) enables targeted interventions:

• Updating factual associations without disturbing related reasoning
• Attenuating sycophancy circuits that over-weight positive-sentiment outputs
• Strengthening refusal circuits without increasing false refusals

Evaluating Value Internalization

Distinguishing a model that has learned a value from one that merely says it has is a core alignment challenge. MI offers a path: reward-hacked models show weaker internal activation of target value features despite matching behavioral scores — a potential ground truth for genuine internalization.

Challenges and Limitations

Scalability remains severe. Full circuit analysis of GPT-2 small (117M parameters) required months of expert time. Scaling to 400B+ parameter models is intractable with current methods.

Feature completeness is unverified. 34M SAE features on Claude 3 Sonnet may be only a fraction of the true feature set.

Causal vs. correlational findings. Many circuit analyses show correlation, not causal necessity. Ablation studies help but don't scale easily.

The moving target problem. Circuit knowledge of Claude 3 Sonnet may not transfer to Claude 3.5 Opus. MI must be applied continuously, not once.

Future Directions

• Automated circuit discovery via LLM-assisted interpretability pipelines (Anthropic experiments, 2025)
• Causal scrubbing at scale (Chan et al., 2022 methodology) via automated tooling
• Cross-model universality mapping — a shared vocabulary of "neural algorithms" transferable between architectures
• Real-time interpretability monitoring as operational safety infrastructure

Dedicated MI teams at Anthropic, DeepMind, EleutherAI, Goodfire, and Transluce signal the field is transitioning from academic curiosity to recognized safety infrastructure.

Conclusion

Mechanistic interpretability is barely begun relative to the complexity of frontier models — but its core hypothesis has accumulated substantial empirical support. Circuits are real. Superposition is real. Feature-level behavioral control is real, if crude.

Without interpretability, alignment depends entirely on behavioral evaluation: observing outputs in test conditions and hoping they generalize. With interpretability, we access a computational ground truth — the difference between a black box and an engine we can inspect.

We are not there yet. But the tools are improving faster than the models are obscuring themselves. That is enough to be cautiously optimistic.

References: Elhage et al. (2022) "Toy Models of Superposition"; Wang et al. (2022) "IOI Circuit in GPT-2"; Olsson et al. (2022) "In-context Learning and Induction Heads"; Zou et al. (2023) "Representation Engineering"; Meng et al. (2022) "ROME"; Chan et al. (2022) "Causal Scrubbing"; Anthropic (2024) "Scaling Sparse Autoencoders"; Anthropic (2025) "On the Biology of a Large Language Model"